The Use of Machine Learning to Predict Hospitalization of Covid-19: A

Case Study in the State of Minas Gerais - Brazil

Gerda Graciela Rodrigues de Oliveira

and Cristiane Neri Nobre

Department of Computer Science, Pontiﬁcal Catholic University of Minas Gerais, PUC Minas - 500, Dom Jos

e Gaspar

Street, Corac¸

ao Eucar

ıstico, Belo Horizonte, Brazil

Keywords:

COVID-19, Hospitalization, Machine Learning.

Abstract:

This work aims to verify the applicability of using Machine Learning techniques to predict hospitalization in

conﬁrmed cases of Covid-19. The study also intends to discover which attributes have the most signiﬁcant

impacts on hospitalization. The machine learning (ML) algorithms used in this experiment were Decision

Tree, Random Forest, Neural Networks, and Naive Bayes. The data used for this experiment were made

available by the government of Minas Gerais - Brazil, through open data. The model based on Random Forest

obtained the best results, presenting the following metrics: Precision, Recall and F1-Score of 0.85, 0.84 and

0.84, respectively. In this experiment, essential characteristics for classifying the patient’s hospitalization are

Comorbidity, Age Group, and HDI. The results point to a good predictive ability, demonstrating the potential

use of ML techniques to predict the hospitalization of people by COVID-19.

1 INTRODUCTION

The year 2020 was marked as a pandemic year. A

Pandemic is a term that signals an epidemiological

trend. Indicates that many outbreaks are happening

simultaneously and spread far and wide. The Coro-

navirus Disease 2019 (COVID-19) has hit everyone

indiscriminately. As a result of the pandemic, many

people were infected, sick, hospitalized, and admitted

to ICU, and many died.

Given the rapid transmission of the virus, there

has been an abrupt and growing additional demand

for hospitalizations worldwide, putting health sys-

tems under strain in many countries (Armstrong et al.,

2020). Brazil has the Uniﬁed Health System (SUS)

which is the largest public and universal health system

in the world, covering the entire country. About 75%

of the Brazilian population does not have a private

https://orcid.org/0000-0002-3521-2367

https://orcid.org/0000-0001-8517-9852

The Uniﬁed Health System (SUS) is one of the largest

and most complex public health systems in the world, rang-

ing from simple attendance to blood pressure assessment,

through from Primary Care to organ transplantation, ensur-

ing complete, universal, and free access to the entire pop-

ulation of Brazil. The SUS was created in 1988 by the

Brazilian Federal Constitution, which determines that the

State must guarantee health to the whole of the Brazilian

population.

health plan and depends exclusively on SUS (Paim

et al., 2011). In this context, the SUS was not pre-

pared to meet the demand. As a result, the country

went through a situation of the imminent collapse of

its health system due to the lack of beds in Intensive

Care Unit (ICU), equipment, essential supplies, and

health professionals to ensure adequate care for pa-

tients in severe conditions affected by the virus (San-

tos et al., 2022b).

The great concern about contracting the disease

revolves around the severity of the disease, which

is variable, with some patients presenting few or no

symptoms, some requiring hospitalization and hospi-

talization in the short, medium, or even long term, di-

rectly affecting public health. Given the above, the

big question is how to predict whether an infected per-

son will be hospitalized.

Machine Learning Models (ML) have been widely

used in health. Simply put, ML is an area of Artiﬁ-

cial Intelligence that attributes the ability to learn au-

tomatically to computational models. Thus, the main

objective of its use in health is to use previous experi-

ences to identify patterns and provide more assertive

information for decision-making.

Given the scenario described and from the

database of conﬁrmed cases of COVID-19, made

available by the government of Minas Gerais-Brazil,

this work has as its primary objective to verify the

392

Rodrigues de Oliveira, G. and Nobre, C.

The Use of Machine Learning to Predict Hospitalization of Covid-19: A Case Study in the State of Minas Gerais - Brazil.

DOI: 10.5220/0011696000003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 392-399

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

applicability of using ML techniques to predict the

hospitalization of patients conﬁrmed with Covid-19,

as well as to present the most signiﬁcant attributes

in this prediction. The machine learning (ML) algo-

rithms used in this experiment were Decision Tree,

Random Forest, Neural Networks, and Naive Bayes.

These algorithms were chosen because they are the

most commonly used for data classiﬁcation.

The remainder of this article is organized as fol-

lows: Section 2 brings the related works; the material

and methods applied in the experiment are described

in Section 3; the last sections get the analysis of the

results, discuss and conclude the ﬁndings, indicating

limitations of the work and proposing opportunities

for further research.

2 RELATED WORKS

Jehi et al. (2020) proposed implementing a predictive

algorithm to estimate, among COVID-19 patients, the

risk of being admitted to a hospital with SARS-CoV-

2 infection based on sex, age, COVID-19 symptoms,

and comorbidities. Algorithm validation was evalu-

ated internally using resampling bootstrap (1000 rep-

etitions) to assess model discrimination and calibra-

tion. Each patient was stratiﬁed as a high, medium,

or low-risk patient for the combined outcome of hos-

pitalization and death, according to clinical and de-

mographic characteristics highlighted by two predic-

tive models developed by ATS Milan: the predictive

model for hospitalization and the predictive model for

overall mortality risk.

Gao et al. (2022) proposed MedML, a machine

learning-based framework that aims to predict hos-

pitalization and severity risk for pediatric patients

with COVID-19 using electronic health records. It

is a hybrid approach that combines medical knowl-

edge graphs and data-driven feature extraction to pre-

dict better pediatric hospitalization and the severity

of COVID-19 MedML extracts the most predictive

attributes based on medical knowledge and propen-

sity scores of more than 6 million medical concepts

and incorporates the relationships between resources

in graphs of medical knowledge through neural net-

works. As a result, MedML achieves an AUROC

to 7% higher and an AUPRC

14% higher compared

to the best baseline machine learning models.

Murtas et al. (2021) aimed to develop a stratiﬁ-

cation tool to improve the management of patients

with COVID-19 and the organization of health care.

AUROC: area under the receiver operating character-

istic curve.

AUPRC: area under the curve of Precision and Recall.

To this end, a predictive algorithm was developed

and applied to 36,834 patients with COVID-19 in

Italy between March 8 and October 9, 2020, to pre-

dict their risk of hospitalization. The algorithm was

developed following TRIPOD ( Transparent report-

ing of a Multivariable Prediction Model for Individ-

ual Prognosis or diagnosis). Exposures considered

were age, sex, comorbidities, and symptoms associ-

ated with COVID-19 (e.g., vomiting, cough, fever, di-

arrhea, myalgia, asthenia, headache, anosmia, ageu-

sia, and dyspnea). As a result, the predictive model

showed an excellent ﬁt to predict hospitalization for

COVID-19 and a good overall prediction accuracy

score of Brier

(0.14).

3 MATERIALS AND METHODS

In this section, we will present the method adopted

to predict the hospitalization of patients conﬁrmed

by Covid-19 in a Brazilian state. The methodology

involves selecting and extracting the database, pre-

processing the dataset, and evaluating metrics. Figure

2 illustrates the steps of the methodology adopted in

this work.

Figure 1: Adopted methodology.

3.1 Obtaining the Database

The database used is available in the form of open

data by the government of Minas Gerais - Brazil.

According to the data dictionary, Table 1, the data

Score Brier is a function that measures the accuracy of

probabilistic predictions. The lower the Brier score is for a

set of predictions, the better the predictions. Its equation is

presented as follows:

BS =

∑

i=1

(pp

− po

)

where n = the number of prediction-observation pairs; pp =

predicted probability of occurrence of the event, and so is

the observed probability of occurrence of the event (zero,

the event did not occur, and one the event did happen).

The Use of Machine Learning to Predict Hospitalization of Covid-19: A Case Study in the State of Minas Gerais - Brazil

393

extraction was performed from the ofﬁcial govern-

ment systems SIVEP Grip (Severe Acute Respira-

tory Syndrome Notiﬁcation System (SARS)), Esus-ve

(tool for recording the notiﬁcation of suspected cases

of Covid-19) and REDCAP (RESEARCH ELEC-

TRONIC DATA CAPTURE, a sophisticated platform

for collecting, managing and disseminating research

data). Data were updated until June 8, 2022, and ex-

tracted from the site

on the same date.

The understanding of the base was carried out

based on the available data dictionary, as well as a

qualitative or quantitative analysis of the base. Two

of the three available spreadsheets were used for this

work. The ﬁrst contains 500,001 records, and the sec-

ond includes 88,101 records, thus totaling 588,102

records and 21 attributes.

3.2 Database Preprocessing

The pre-processing of the database can be deﬁned as

a set of activities that involve the preparation, orga-

nization, and structuring of the data. It is an essential

step that precedes the accomplishment of the analyses

and predictions since it will be decisive for the qual-

ity of the data analyzed. The following subsections

describe each pre-processing step performed.

3.2.1 Selection of Attributes

Attribute selection performs a primary task and

aims to select the more important attributes since

non-relevant and/or redundant features can re-

duce the accuracy used by machine learning algo-

rithms. For this work, the selection was made

from understanding the data contained in Table

1. Of the 21 constant attributes, seven features

and one class were selected. The selected at-

tributes were: MICRO, DATE NOTIFICATION,

SEX, AGE, DATE 1st SYMPTOM, RACE, CO-

MORBIDITY, and the HOSPITALIZATION class.

3.2.2 Data Cleanup

The database may contain irrelevant or missing data.

To deal with this situation, it is necessary to perform

data cleaning. Both researchers and practitioners have

recognized the need to detect errors, repair, and rea-

son about querying and analyzing dirty and inconsis-

tent data (Ilyas and Rekatsinas, 2022). In the case

of the data presented, the instances containing miss-

ing data or “Not informed” conditions were removed,

and the removal of inconsistent values were. A total

of 112,917 instances were excluded in this process, so

https://coronavirus.saude.mg.gov.br/

after cleaning the data, the base was left with 475,185

records.

3.2.3 Transformation, Normalization, and

Discretization of Data

This step transforms the original data into more ap-

propriate and suitable formats for machine learning

algorithms. This is important, as some algorithms

only work with numerical value entries, unable to pre-

dict others types of data entries. At this stage, the

URS attribute was discarded, and the micro attribute

was used because it had a more signiﬁcant number

of responses (89 different responses). This attribute

MICRO (Microregion) has been transformed into two

new attributes, namely: Population and HDI ( Human

Development Index) footnote The Human Develop-

ment Index (HDI) is a general and synthetic measure

used to classify the degree of economic development

and the quality of life of countries. It was created

in 1990 and published annually since 1993 by the

United Nations Development Program – UNDP. The

HDI varies on a scale from 0 to 1. The closer to 1,

the greater the human development.. This informa-

tion was collected from data provided by the IBGE

(Brazilian Institute of Geography and Statistics). For

the HDI, there was no need to take any action since it

is already normalized data.

The ‘POPULATION’ attribute was presented in

absolute value according to each microregion. The

same was discretized in 3 bands: small, medium, and

large. For the small population, a value of fewer

than 50 thousand inhabitants was considered; for the

medium-sized population, values equal to or greater

than 50,000 inhabitants and less than 100,000 inhabi-

tants were considered; for the large population, a pop-

ulation similar to or greater than 100,000 inhabitants

was considered.

The “GENDER” attribute was coded as Female =1

and Male =0.

The RACE attribute was binarized and trans-

formed into ﬁve columns, white, brown, black, yel-

low, and indigenous, thus representing the presence

or absence of the characteristic.

The AGE attribute was discretized as follows:

Children = Under 12 years old, Adolescents = 12 to

18 years old

, Young = 19 to 29 years old

, Adults =

These transformations carried out for children and

adolescents were created according to the ECA (Child and

Adolescent Statute). The ECA was made through Federal

Law n

8.069, July 13, 1990, which regulates article 227 of

the Federal Constitution of 1988.

The transformation to “youth” is following the Youth

Statute. This statute was created through Federal Law No.

12,852, of August 5, 2013.

HEALTHINF 2023 - 16th International Conference on Health Informatics

394

Table 1: Data dictionary of the database provided by the government of a Minas Gerais - Brazil.

ATTRIBUTE DESCRIPTION

ID Unique exam code

URS Name of the URS (Regional Health Unit) where the test was performed. There is the possibility of 28 different

answers regarding the URS of the Brazilian State considered

MICRO Name of the Microregion in which the exam was performed. There is the possibility of 89 different answers

regarding the Microregions of the Brazilian State considered

MACRO Name of the Macro-region where the exam was performed. There is the possibility of 14 different answers

regarding the Macroregions of the Brazilian State considered

NOTIFICATION DATE Date of notiﬁcation of the patient’s Exam result

CLASSIFICATION CASE Result of the exam performed by the patient (if is a conﬁrmed case for COVID-19)

GENDER Gender of the patient who underwent the examination (male, female or not informed)

AGE Age of the patient who took the exam

AGE GROUP Patient’s age group

CITIES RESIDENCE Name of the patient’s municipality of residence. In this ﬁeld, there is the possibility of 853 different answers

CODE IBGE code related to the patient’s municipality of residence. There is the possibility of 853 different answers.

EVOLUTION Evolution of the patient’s condition

EVOLUTION DATE Date of evolution of the patient’s condition

DATE 1 SYMPTOM Date of 1st symptoms of the case.

HOSPITALIZATION Situation of hospitalization in the patient’s clinical bed

UTI Situation of the patient’s ICU bed admission

RACE Patient’s race

ETHNICITIES Names of indigenous ethnicities. Name and ethnicity code of the patient, when indigenous

COMORBIDITY Presence or absence of patient comorbidity

DATE UPDATE Data regarding update data Status

SOURCE INFORMATION Laboratory that performed the examination of the patient

30 to 59 years old, Elderly = 60 years old or older

For the attributes “NOTIFICATION DATE” and

“DATA 1

SINTOMA” two new columns were cre-

ated, being the month and year.

The COMORBIDITY attribute has the answers

“Yes”, “No”, and “Not informed” (considered miss-

ing data). thus was binarized as follows: Yes =1 and

No =0.

An attribute called “ILLNESS TIME” was cre-

ated, which is the calculation of the difference be-

tween the attributes “NOTIFICATION DATE” and

“DATE 1

SYMPTOM”. Negative values or values

greater than 150 days were considered inconsistent

and were removed from the database. This time was

normalized using the Min-Max method.

According to Ciaburro et al. (2018), min-max nor-

malization performs a linear transformation on the

original data. This technique gets all data scaled in

the range (0, 1), Equation 1.

Scaled

X − X

min

max

− X

min

(1)

Also, according to Ciaburro et al. (2018), the min-

max normalization preserves the relationships be-

tween the original data values. The cost of limiting

this range is that we will end up with more minor

standard deviations, which can suppress the effect of

outliers.

According to the Elderly Person Statute, it was created

by Federal Law No. 10,741 of October 1, 2003.

Thus, after these pre-processing steps, the base

presented 475,185 instances, 15 input attributes,

and 1 Class attribute. The attributes are POPU-

LATION, HDI, DATE-NOT-MONTH, DATE-NOT-

YEAR, GENDER, AGE GROUP, DATE-1-SINT-

MONTH, DATE-1-SINT-YEAR, ILLNESS TIME,

COMORBIDITY, BROWN RACE, RACE WHITE,

YELLOW RACE, BLACK RACE, INDIGENOUS

RACE, and HOSPITALIZATION class.

3.2.4 Redundancy Elimination

Redundancy in machine learning preprocessing is the

presence of attributes that are highly correlated with

each other. Removing redundancy is helpful because

it can reduce training time and increase model accu-

racy. Furthermore, eliminating redundancy can sim-

plify the model, making it easier to interpret.

The drop duplicates() function was used to ex-

clude redundant data from the database, which elimi-

nates equal instances from the database. In this way,

the unique values in the base were kept; that is, there

was no data loss. In this process, we went from

475,185 instances to 185,742; approximately 61% of

the database corresponds to repeated cases.

3.2.5 Missing Data Imputation

After removing the redundancy, missing data (null)

was evaluated. There are different methods of imput-

ing missing data. In this case, the KNNImputer was

The Use of Machine Learning to Predict Hospitalization of Covid-19: A Case Study in the State of Minas Gerais - Brazil

395

used, which assigns the missing values using the Eu-

clidean distance. KNNimputer is a scikit -learn class

used to ﬁll in or predict missing values in a dataset.

The n neighs =2 (number of data points to be included

closer to the missing value) was used as a parameter

for this process.

3.3 Description of Methods

The study was conducted according to the workﬂow

shown in Figure 2. To obtain the best predictive ca-

pacity, models based on four ML algorithms were

used: Decision Tree, Random Forest, Neural Net-

works, and Naive Bayes. The models were imple-

mented in Python using the scikit-learn learning li-

brary, developed explicitly for the practical applica-

tion of Ml. After some tests, we present the hyperpa-

rameters used in the experiment. The system’s default

hyperparameters were utilized for the Naive Bayes

method.

The following hyperparameters were used for the

Decision Tree: Criterion = ‘entropy’ (the criterion

for measuring the quality of a division); Max depth

= 5 (the maximum depth of the tree); Max features

= ‘auto’ (the maximum number of features to be con-

sidered for the best split); Min samples leaf = 7 (the

minimum number of samples needed to be in a leaf

node); Min samples split = 6 (the minimum number

of the samples required to split a node); Splitter =

‘random’ (the criterion used to choose the best divi-

sion).

For Neural Networks, the following hyperparam-

eters were used: Activation = ‘relu’ (Activation func-

tion for hidden layer); Alpha = 0.001 (Regularization

Term); Hidden layer sizes = 20 (Number of neurons

in the hidden layer); Learning rate init = 0.001 (Ini-

tial learning rate); Max iter = 700 (Maximum num-

ber of iterations); Solver = ‘adam’ (algorithm for op-

timization).

For Random Forest, the following hyperparame-

ters were used: n estimators= 100 (the number of

trees); Criterion = ‘entropy’, (the criterion used to

measure the quality of a node’s split); Max depth =

10 (the maximum depth of the tree); Max features =

10 (the number of features to consider when looking

for the best split); n jobs=5 (the number of cores to

be used to build the trees).

3.4 Model Quality Assessment Metrics

The following metrics will be used to measure the

quality of the models: Recall

, Precision

, and F-

Recall =

T P

T P+F N

Precision =

T P

T P+F P

measure

For the evaluation of the quality of the learn-

ing models, 10% of the data was reserved for test-

ing. In the remaining 90% of the data, 10-fold cross-

validation was applied. Figure 2 presents a summary

of the entire process performed.

4 RESULTS AND DISCUSSIONS

The coronavirus is one of the pathogens that act on

the human respiratory system, causing severe acute

respiratory syndrome (SARS) and Middle East res-

piratory syndrome (MERS), and it is considered a

pathogen of great importance to public health (Rothan

and Byrareddy, 2020). Since the ﬁrst case of SARS-

CoV-2 infection, modeling the epidemic growth pat-

tern has been crucial to understanding the pandemic’s

evolution and guiding the implementation of preven-

tion and control measures (Sciannameo et al., 2022).

The outbreak caused by the coronavirus was de-

clared in Brazil a pandemic, on March 11, 2020,

by the World Health Organization (WHO) and, thus,

brought signiﬁcant challenges to public health. In

Brazil, the ﬁrst case of COVID-19 was conﬁrmed on

February 26, and the ﬁrst death on March 17. Fur-

thermore, the Brazilian Ministry of Health recognized

national community transmission on March 20, 2020

(Santos et al., 2022a).

Regarding clinical manifestations, the Epidemio-

logical Surveillance Guide of the Brazilian Ministry

of Health (da S

aude do Brasil, 2021) clariﬁes that

SARS-CoV-2 infection can range from asymptomatic

cases and mild clinical manifestations to moderate,

severe, and critical conditions. It also highlights that

although most people with covid-19 develop mild

(40%) or medium (40%) symptoms, approximately

15% may develop severe symptoms that require oxy-

gen support, and about 5% may have the necessary

form of the disease. In this situation, the individ-

ual may experience complications such as respiratory

failure, sepsis and septic shock, thromboembolism,

and multiple organ failure, including acute liver or

heart damage, requiring intensive care. These same

rates are presented by the World Health Organization

(Organization et al., 2020).

According to the Hospital Policy and Manage-

ment Observatory (OPGH) (Albuquerque, 2021), in

2020, there was an increase of almost 55% compared

to 2019 in hospitalization for treatment of infectious

and parasitic diseases (IPD) in the SUS and was the

ﬁrst cause of hospitalization in 2020.

F-measure = 2 ×

Precision×Recall

Precision+Recall

HEALTHINF 2023 - 16th International Conference on Health Informatics

396

Figure 2: Methodology used in data processing.

At different times, the COVID-19 pandemic in

Brazil led the health system to exhaust its installed

capacity. Thus, it became necessary to capture the ex-

tent of the damage and build knowledge to subsidize

care reorganization choices to deal with unmet routine

needs and the new demands imposed by Covid. So, it

is essential to trace the pattern of use in the Uniﬁed

Health System (SUS) hospital network during this

pandemic(Portela et al., 2021). Analyzing this pattern

will enable more assertive decision-making regarding

the patient’s hospitalization.

For the analyzed dataset, we executed the mod-

els of the ﬁve proposed ML methods (Decision Tree,

Random Forest, Neural Networks, and Naive Bayes)

to understand the proﬁle of people hospitalized for

COVID-19. Table 2 shows the demographic data of

the study.

Table 2: Demographic questions.

Question N %

Gender

Male 93,042 50.09

Female 92,700 49.91

Age group

12 less 10,467 5.64

12 to 18 9,961 5.36

19 to 29 27,656 14.89

30 to 59 74,121 39.91

60 or more 63,537 34.21

Race

White 80,335 43.25

Brown 76,385 41.12

Black 20,985 11.30

Yellow 7,725 4.16

Indigenous 312 0.17

Population

50k less 31,879 17.16

50k to 99.999 52,052 28.02

100k or more 101,811 54.81

Observing Figure 3, which presents the results of

the testing phase of the prediction models, the model

based on Random Forest obtained the best results,

considering that the model has an excellent predictive

capacity for this dataset. Thus, we used the Random

Forest method to discuss the results better.

Figure 3: Comparison of performance in tests of prediction

models.

After choosing the model, it is necessary to un-

derstand the quality of its results and make the model

results more explainable. To this end, the importance

of each attribute selected by the model for the predic-

tion was veriﬁed, where the most critical feature was

comorbidity, followed by age group.

The importance of each attribute was calculated

using Random Forest’s feature importances method.

This method is applied to Random Forest models and

returns a list containing the importance of each fea-

ture for the model. The importance of attributes is

calculated according to the average impurity reduc-

tion when a decision vertex is divided on an element.

The importance of the attributes was displayed

through a bar graph (Figure 4), where the length of

the rectangle represents the weight of the feature.

Figure 4: Importance of attributes.

This graph represents the weight of each attribute

for the classiﬁcation of a patient’s hospitalization.

Thus, in this experiment, essential characteristics for

classifying the patient’s hospitalization are: Comor-

The Use of Machine Learning to Predict Hospitalization of Covid-19: A Case Study in the State of Minas Gerais - Brazil

397

bidity, Age Group, and HDI.

Feitoza et al. (2020) state that about 20% to

51% of patients with COVID-19 were detected with

a chronic disease already installed. Ranzani et al.

(2021) shows in its study that a large proportion, 74%,

of hospitalized patients with COVID-19 in Brazil had

one or two comorbidities. Regarding the Age Group,

it can be seen in Figure 5, taken from the database,

that the elderly and adults have a greater tendency

to be hospitalized by Covid-19 than children, adoles-

cents, and young people.

Figure 5: Age group by class.

Lana et al. (2021) presented in tgeir study the

over-risk calculation and the incidence rate of hos-

pitalization and deaths per thousand inhabitants for

Covid-19 by gender, age group, and comorbidity. The

study shows that from the age group of 45-49 years

(SRfe

= 1.133), there is a gradual increase in the

risk of hospitalization, reaching SRfe = 8.466 among

those aged 90 years or older, which demonstrates that

age is a strong and well-established risk factor for

hospitalization and death from COVID-19.

Regarding the HDI Maciel et al. (2020) analyzed

the spatial distribution of the incidence of Covid-19

and its correlation with the municipal human devel-

opment index (MHDI) of municipalities in a Brazilian

state. They concluded that the incidence of COVID-

19 was heterogeneous and unequally distributed for

the Brazilian state studied and that the cities with the

highest coefﬁcients of incidence of COVID-19 had

higher MHDI values. Thus, HDI can be presented

as a relevant factor for the patient with Covid-19.

The graph represented by Figure 6, drawn from

the data of this study, collaborates with the ﬁndings

above and shows the increase in hospitalization and

the rise to the increase in HDI.

SRfe = The stratiﬁed over-risk is deﬁned for group i

with the denominator of the incidence rate for the age group

(fe) in question, that is, SR

( f e) =

, f e

, where Θ

, f e and

, f e represent the incidence rate of group i and the gen-

eral population for the age group of interest. The SRE

( f e)

distribution for group i has derived analogously to the over-

risk distribution for group i.

Figure 6: HDI by class.

The analysis showed a positive correlation be-

tween the incidence of hospitalization for Covid-19

and the HDI; that is, increases in the HDI also in-

crease the incidence of hospitalization for Covid-19.

5 FINAL CONSIDERATIONS

This work demonstrates the potential use of ML tech-

niques to predict hospitalizations for COVID-19. The

results presented in this study are considered promis-

ing since the ML models performed well in classify-

ing patients. In addition, ML models showed results

on time, which is extremely important regarding re-

sponse time. Therefore, ML models can be used as

additional tools in decision-making regarding patients

hospitalized for COVID-19.

Implementing an ML model can provide a more

accurate assessment of the hospitalization of a pa-

tient with COVID-19 and assist the health agent in

decision-making. If systematically implemented, and

with the breadth of clinical data collected, it is be-

lieved that better identiﬁcation and care of patients

infected by COVID-19 will be possible.

As a suggestion for future work, the ML models

presented in this study can be improved with other

ML hyperparameters and/or methods, such as the

Deep Neural Network, which gives a better represen-

tation of data and, therefore, can help to improve the

performance of the model. In addition, other open

datasets can also be analyzed to understand better the

proﬁle of patients hospitalized for COVID-19.

5.1 Study Limitation

Knowing that the clinical spectrum of COVID-19 is

vast, the a priori clinical knowledge about the asso-

ciations between symptoms and hospitalization for

COVID-19 was limited for this study since the ﬁeld

“Comorbidity” presented only the yes or no values.

Thus, we can point out that the major limitation of

this research is the lack of data for an evaluation of

other important variables in this predictive process,

such as the patient’s clinical condition, diabetes, high

HEALTHINF 2023 - 16th International Conference on Health Informatics

398

blood pressure, chronic kidney disease, fever, cough,

shortness of breath, fatigue, lung disease, among oth-

ers.

ACKNOWLEDGEMENTS

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior –

Brasil (CAPES) – Finance Code 001. The authors

thank the National Council for Scientiﬁc and Techno-

logical Development of Brazil (CNPq - Conselho Na-

cional de Desenvolvimento Cient

ıﬁco e Tecnol

ogico)

and the Foundation for Research Support of the Minas

Gerais State (FAPEMIG). The work was developed

at the Pontiﬁcal Catholic University of Minas Gerais,

PUC Minas.

REFERENCES

Albuquerque, C. (2021). Pandemia diminui n

umero e muda

perﬁl de internac¸

oes no sus em 2020. Dispon

ıvel

em: https://observatoriohospitalar.ﬁocruz.br/debates-

e-opinioes/pandemia-diminui-numero-e-muda-perﬁl-

de-internacoes-no-sus-em-2020. Acesso em:01 de

julho de 2022.

Armstrong, R., Kane, A., and Cook, T. (2020). Out-

comes from intensive care in patients with covid-19: a

systematic review and meta-analysis of observational

studies. Anaesthesia, 75(10):1340–1349.

Ciaburro, G., Ayyadevara, V. K., and Perrier, A. (2018).

Hands-on machine learning on google cloud plat-

form: Implementing smart and efﬁcient analytics us-

ing cloud ml engine. Packt Publishing Ltd.

da S

aude do Brasil, M. (2021). Guia de vigil

ancia

epidemiol

ogica para covid-19. Dispon

ıvel em:

https://coronavirus.saude.mg.gov.br/images/1

2021/17-

03-Guia de vigilancia da covid 16marc2021.pdf.

Acesso em:01 de julho de 2022.

Feitoza, T. M. O., Chaves, A. M., Muniz, G. T. S., da Cruz,

M. C. C., and Junior, I. d. F. C. (2020). Comorbidades

e covid-19. Revista Interfaces: Sa

ude, Humanas e

Tecnologia, 8(3):711–723.

Gao, J., Yang, C., Heintz, J., Barrows, S., Albers, E.,

Stapel, M., Warﬁeld, S., Cross, A., Sun, J., et al.

(2022). Medml: Fusing medical knowledge and

machine learning models for early pediatric covid-

19 hospitalization and severity prediction. Iscience,

25(9):104970.

Ilyas, I. F. and Rekatsinas, T. (2022). Machine learning and

data cleaning: Which serves the other? ACM Journal

of Data and Information Quality (JDIQ), 14(3):1–11.

Jehi, L., Ji, X., Milinovich, A., Erzurum, S., Merlino, A.,

Gordon, S., Young, J. B., and Kattan, M. W. (2020).

Development and validation of a model for individual-

ized prediction of hospitalization risk in 4,536 patients

with covid-19. PloS one, 15(8):e0237419.

Lana, R. M., Freitas, L. P., Codec¸o, C. T., Pacheco, A. G.,

Carvalho, L. M. F. d., Villela, D. A. M., Coelho, F. C.,

Cruz, O. G., Niquini, R. P., Porto, V. B. G., et al.

(2021). Identiﬁcac¸

ao de grupos priorit

arios para a

vacinac¸

ao contra covid-19 no brasil. Cadernos de

ude P

ublica, 37.

Maciel, J. A. C., Castro-Silva, I. I., and Farias, M. R. d.

(2020). An

alise inicial da correlac¸

ao espacial entre a

incid

encia de covid-19 e o desenvolvimento humano

nos munic

ıpios do estado do cear

a no brasil. Revista

brasileira de epidemiologia, 23:e200057.

Murtas, R., Morici, N., Cogliati, C., Puoti, M., Omazzi, B.,

Bergamaschi, W., Voza, A., Querini, P. R., Stefanini,

G., Manfredi, M. G., et al. (2021). Algorithm for indi-

vidual prediction of covid-19–related hospitalization

based on symptoms: Development and implementa-

tion study. JMIR Public Health and Surveillance,

7(11):e29504.

Organization, W. H. et al. (2020). Oxygen sources and dis-

tribution for covid-19 treatment centres: interim guid-

ance, 4 april 2020. Technical report, World Health

Organization.

Paim, J., Travassos, C., Almeida, C., Bahia, L., and

Macinko, J. (2011). The brazilian health system:

history, advances, and challenges. The Lancet,

377(9779):1778–1797.

Portela, M. C., de Aguiar Pereira, C. C., Lima, S. M. L.,

de Andrade, C. L. T., and Martins, M. (2021). Patterns

of hospital utilization in the uniﬁed health system in

six brazilian capitals: comparison between the year

before and the ﬁrst six ﬁrst months of the covid-19

pandemic. BMC Health Services Research, 21(1):1–

13.

Ranzani, O. T., Bastos, L. S., Gelli, J. G. M., Marchesi,

J. F., Bai

ao, F., Hamacher, S., and Bozza, F. A. (2021).

Characterisation of the ﬁrst 250 000 hospital admis-

sions for covid-19 in brazil: a retrospective analysis

of nationwide data. The Lancet Respiratory Medicine,

9(4):407–418.

Rothan, H. A. and Byrareddy, S. N. (2020). The epidemi-

ology and pathogenesis of coronavirus disease (covid-

19) outbreak. Journal of autoimmunity, 109:102433.

Santos, L. G., da Silva, R. V., Leal, T. C., Xavier, J. E.,

de Souza Figueiredo, E. V. M., de Paiva, J. P. S.,

da Silva, L. F., de Oliveira Rocha, C. A., Nunes, B. E.

B. R., de Almeida Santana, G. B., et al. (2022a). Im-

pact of the covid-19 pandemic on hospital admissions

and in-hospital lethality from cardiovascular diseases

in brazil: an ecological and time series study covid-

19 and cardiovascular diseases. Current Problems in

Cardiology, page 101216.

Santos, P. P. G. V. d., Oliveira, R. A. D. d., and Albuquerque,

M. V. d. (2022b). Desigualdades da oferta hospitalar

no contexto da pandemia da covid-19 no brasil: uma

revis

ao integrativa. Sa

ude em Debate, 46:322–337.

Sciannameo, V., Gofﬁ, A., Maffeis, G., Gianfreda, R.,

Pagliari, D. J., Filippini, T., Mancuso, P., Giorgi-

Rossi, P., Dal Zovo, L. A., Corbari, A., et al. (2022).

A deep learning approach for spatio-temporal fore-

casting of new cases and new hospital admissions of

covid-19 spread in reggio emilia, northern italy. Jour-

nal of Biomedical Informatics, 132:104132.

The Use of Machine Learning to Predict Hospitalization of Covid-19: A Case Study in the State of Minas Gerais - Brazil

399